Research map / Probabilistic and scientific modeling / Molecules and proteins
Full-Atom Peptide Design with Geometric Latent Diffusion
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes PepGLAD, a full-atom latent diffusion model for receptor-conditioned peptide design that uses a VAE over variable-size residues plus a receptor-specific affine transformation to a standard space, and releases the PepBench benchmark.
Problem
Target-specific peptide design given a binding site: generating peptide sequence and full-atom 3D structure, which existing methods handle inefficiently or only as target-agnostic 1D sequences, and diffusion models generalize poorly across variable binding-site geometries.
Approach
Runs diffusion in the latent space of a full-atom VAE (variable-size residues encoded to fixed-dimensional latents) and applies a receptor-specific affine transformation (center offset plus Cholesky of binding-site covariance) to map geometry to a standard Gaussian space, while preserving E(3)-equivariance.
Key results
- On PepBench, PepGLAD reaches diversity 0.506, consistency 0.789, and 55.97% success rate versus DiffAb's 0.427, 0.670, and 49.87%.
- In binding conformation generation on PepBench, PepGLAD achieves RMSD Cα 4.09 Å and DockQ 0.592, beating DiffAb (4.23 Å, 0.586).
- Against RFDiffusion, PepGLAD reaches 92.82% success at about 3s per candidate versus 61s for RFDiffusion, though RFDiffusion has better ΔG (-33.82 vs -29.36).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use PepGLAD instead of DiffAb | PepBench sequence-structure co-design · PepBench | diversity: 0.427 -> 0.506 | Helped |
| Use PepGLAD instead of DiffAb | PepBench sequence-structure co-design · PepBench | consistency: 0.670 -> 0.789 | Helped |
| Use PepGLAD instead of DiffAb | PepBench sequence-structure co-design · PepBench | ΔG: -21.20 -> -21.94 | Helped |
| Use PepGLAD instead of DiffAb | PepBench sequence-structure co-design · PepBench | success: 49.87% -> 55.97% | Helped |
| Use PepGLAD instead of RFDiffusion | sequence-structure co-design · Sequence-structure co-design | success: 79.68% -> 92.82% | Helped |
| Use PepGLAD instead of DiffAb | PepBench binding conformation generation · PepBench | RMSD atom: 7.60 -> 5.30 | Helped |
| Add full-atom geometry (ablation) | PepGLAD sequence-structure co-design · PepBench | success: 51.18% -> 55.97% | Helped |
| Add receptor-specific affine transformation (ablation) | PepGLAD sequence-structure co-design · PepBench | average normalized metrics: 0.564 -> 0.619 | Helped |
| Add unsupervised ProtFrag data to autoencoder training (ablation) | PepGLAD sequence-structure co-design · PepBench | success: 52.15% -> 55.97% | Helped |
| Add mask policy to autoencoder training (ablation) | PepGLAD sequence-structure co-design · PepBench | consistency: 0.741 -> 0.789 | Helped |
Surprising findings
- Amino acid recovery is nearly uninformative for target-specific peptide design; models with similar AAR (e.g. HSRN 35.8%, DiffAb 37.1%, PepGLAD 36.7%) differ widely in success rate.
- Even AlphaFold 2 predicts only 36% of test complexes within 5 Å RMSD, so supervised consistency against AF2 structures would be unreliable.
Limitations
- Binding affinity is assessed with Rosetta energy as a proxy rather than wet-lab experiments.
- Peptide design is treated only from the protein perspective; small-molecule design methods are not explored.
- The sequence-order guidance is a minor technical trick with small performance gains.
Experimental setup
Datasets: PepBench ProtFrag PepBDB LNR
Benchmarks: PepBench PepBDB
Metrics: diversity consistency (Cramer's V) binding energy (Rosetta ΔG) success rate RMSD Cα RMSD atom DockQ AAR
Baselines: HSRN dyMEAN DiffAb RFDiffusion AlphaFold 2 AnchorExtension FlexPepDock
Builds on: dyMEAN latent diffusion models DDPM classifier-guided diffusion
Closest papers by meaning
- Full-Atom Peptide Design based on Multi-modal Flow Matching (ICML 2024 · similarity 0.91)
- Generative Modeling of Full-Atom Protein Conformations using Latent Diffusion on Graph Embeddings (NeurIPS 2025 · similarity 0.91)
- UniMoMo: Unified Generative Modeling of 3D Molecules for De Novo Binder Design (ICML 2025 · similarity 0.89)
- Hotspot-Driven Peptide Design via Multi-Fragment Autoregressive Extension (ICLR 2025 · similarity 0.88)
- PepTune: De Novo Generation of Therapeutic Peptides with Multi-Objective-Guided Discrete Diffusion (ICML 2025 · similarity 0.88)
- BindGPT: A Scalable Framework for 3D Molecular Design via Language Modeling and Reinforcement Learning (AAAI 2025 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.